In my work on the remaining capacity forecast of an electric vehicle battery pack, I have focused on the fact that the battery is not merely an energy storage component but the central determinant of vehicle range, acceleration, climbing ability, and overall operating reliability. For an electric vehicle battery pack, the ability to estimate the remaining capacity accurately and quickly is one of the most important functions of a battery management system. I began with the practical observation that the electric vehicle battery pack is a complex, nonlinear, and strongly coupled electrochemical system. Its internal reactions vary with current, voltage, temperature, aging, and previous operating history. Because of these characteristics, a fixed mathematical model is difficult to construct with high fidelity. Therefore, I investigated data-driven forecasting methods, especially neural networks and genetic algorithms, to predict the remaining capacity of a lithium iron phosphate electric vehicle battery pack.
The work I describe here is organized around a clear engineering objective: to predict the remaining capacity of an electric vehicle battery pack from measurable quantities such as discharge current and discharge voltage, and to improve both accuracy and training efficiency by combining a back propagation neural network with a genetic algorithm. I use the term electric vehicle battery pack repeatedly because the method is intended for pack-level and cell-level monitoring, where the state of the electric vehicle battery pack directly influences the usable range and safety margin. The forecast problem can be summarized as estimating the remaining charge from observable terminal behavior. I first studied the discharge behavior of lithium iron phosphate cells, then derived the input-output relationship for a neural network, then built and optimized a BP model, and finally validated the resulting system against independent measurements.
1. Why Remaining Capacity Matters for an Electric Vehicle Battery Pack
For an electric vehicle battery pack, the remaining capacity is not simply a display value. It determines how far the vehicle can travel before recharge, how much power can be delivered during acceleration, and how much regenerative braking energy can be accepted without overvoltage. If the remaining capacity is overestimated, the driver may experience an unexpected shutdown. If it is underestimated, the electric vehicle battery pack is used conservatively, and its available energy is wasted. Therefore, accurate remaining capacity forecasting has both theoretical and practical value for electric vehicle battery pack design, testing, and operation.
The conventional methods I examined include capacity tests, open-circuit voltage methods, internal resistance methods, electrolyte density methods, ampere-hour integration, recovery-effect methods, physicochemical models, fuzzy inference, and neural networks. Each has limitations when applied to an electric vehicle battery pack. A full capacity test is accurate but time-consuming and destructive to the operating schedule. Open-circuit voltage requires a long rest period and is unsuitable for online estimation. Internal resistance is difficult to measure online and does not have a strict one-to-one relationship with remaining capacity. Electrolyte density is impractical for sealed cells. Ampere-hour integration is simple but accumulates error because of self-discharge, temperature, and aging. Physicochemical models can be accurate but require many parameters and heavy computation. These limitations motivated my use of neural-network forecasting for the electric vehicle battery pack.
| Method | Main Principle | Advantage | Limitation for Electric Vehicle Battery Pack |
|---|---|---|---|
| Full capacity test | Constant-current discharge to cutoff | High accuracy | Time-consuming, reduces cycle life |
| Open-circuit voltage | Relation between OCV and state of charge | Simple when battery is rested | Not suitable for online use |
| Internal resistance | Resistance change with state | Sensitive to health | Online measurement is difficult |
| Ampere-hour integration | Integral of current over time | Easy to implement | Error accumulates with aging and temperature |
| Physicochemical model | Electrochemical and thermal equations | Interpretable | Many parameters and high computation |
| Neural network | Nonlinear input-output mapping | Good for uncertain systems | Can fall into local minima if not optimized |
2. Lithium Iron Phosphate Chemistry and Performance Indicators
The electric vehicle battery pack I studied is based on lithium iron phosphate chemistry. In this chemistry, lithium ions move between a negative electrode and a positive electrode during charge and discharge. The positive electrode uses lithium iron phosphate, and the negative electrode typically uses a carbon-based material. The general chemical expression can be written as
$$ \mathrm{Li_xC_6 | LiClO_4 | Li_{1-x}FePO_4} $$
The positive electrode reaction can be represented as
$$ \mathrm{LiFePO_4 \rightleftharpoons Li_{1-x}FePO_4 + xLi^+ + xe^-} $$
The negative electrode reaction can be represented as
$$ \mathrm{xLi^+ + xe^- + 6C \rightleftharpoons Li_xC_6} $$
For an electric vehicle battery pack, this chemistry offers high safety, good thermal stability, a flat discharge plateau, and long cycle life. The nominal cell voltage is approximately 3.2 V, the upper charge voltage is about 3.65 V, and the lower discharge cutoff is about 2.0 V for the cell type I used. The capacity is usually expressed in ampere-hours, and the energy is expressed in watt-hours. The discharge rate is defined as
$$ \text{Discharge Rate} = \frac{I}{C_{\text{rated}}} $$
where \(I\) is the discharge current and \(C_{\text{rated}}\) is the rated capacity. For an electric vehicle battery pack, the discharge rate affects both the terminal voltage and the available capacity. A higher discharge rate generally lowers the voltage plateau and reduces the usable capacity because of polarization and internal resistance.
The capacity delivered by a cell or an electric vehicle battery pack is calculated from current and time:
$$ C = I t $$
If the current varies with time, the delivered capacity is
$$ C_{\text{out}}(t) = \int_0^t I(\tau)\,d\tau $$
The remaining capacity can then be written as
$$ C_{\text{rem}}(t) = C_{\text{rated}} – C_{\text{out}}(t) $$
For an electric vehicle battery pack, the state of charge is often defined as
$$ \mathrm{SOC}(t) = \frac{C_{\text{rem}}(t)}{C_{\text{rated}}} \times 100\% $$
The Peukert relationship is also useful for describing how available capacity changes with current:
$$ I^n T = K $$
or equivalently
$$ C_I = K I^{1-n} $$
where \(K\) and \(n\) are empirical parameters. This equation shows that the electric vehicle battery pack cannot be treated as an ideal linear energy reservoir. The available capacity depends on the discharge current and on the operating conditions.
| Battery Type | Specific Energy (Wh/kg) | Energy Density (Wh/L) | Specific Power (W/kg) | Cycle Life (cycles) |
|---|---|---|---|---|
| VRLA | 35-40 | 60-90 | 200-300 | 400-6000 |
| Ni-Cd | 40-60 | 80-110 | 150-350 | 600-1200 |
| Ni-MH | 60-70 | 130-170 | 150-300 | 300-1200 |
| Na/S | 100 | 150 | 200 | 800 |
| Li-Polymer | 155 | 220 | 315 | 600 |
| Li-Ion | 90-130 | 140-200 | 250-450 | 800-1200 |
3. Discharge Experiments and Data Acquisition
To obtain realistic data for the electric vehicle battery pack, I used a lithium iron phosphate cell with a nominal voltage of 3.2 V and a rated capacity of 3 Ah. I charged the cell to 3.65 V and then rested it for several hours before discharge. I then discharged the cell at different constant currents and recorded voltage, current, and delivered capacity. The sampling interval was fixed by the battery testing equipment. I also measured internal resistance under open-circuit conditions. The experiments were repeated at different currents so that the model could learn the nonlinear relationship between voltage, current, and remaining capacity.
The voltage measurement is not simply a two-terminal measurement when the current is large. For an electric vehicle battery pack, contact resistance and cable resistance can influence the reading. I therefore considered both two-wire and four-wire measurement arrangements. In a simplified form, the terminal voltage under load is
$$ U = E – I R_0 $$
where \(E\) is the electromotive force, \(I\) is the current, and \(R_0\) is the internal resistance. When contact resistance is included, the measured voltage can be expressed as
$$ U_{\text{meas}} = E – I(R_0 + R_{\text{contact}}) $$
Current was measured in series with the cell or through a shunt, and capacity was obtained by integrating current over time. The capacity equation used in the experiment was
$$ C = \sum_{k=1}^{N} I_k \Delta t_k $$
where \(I_k\) is the current at sample \(k\), \(\Delta t_k\) is the sampling interval, and \(N\) is the number of samples. The remaining capacity was then computed as the difference between the rated capacity and the delivered capacity.
| Discharge Current (A) | Terminal Voltage (V) | Delivered Capacity (Ah) |
|---|---|---|
| 3 | 3.2267 | 0.04991 |
| 3 | 3.2142 | 0.29978 |
| 3 | 3.2132 | 0.34967 |
| 3 | 3.2117 | 0.39971 |
| 3 | 3.2015 | 0.69943 |
| 4 | 3.0848 | 1.2242 |
| 4 | 3.0817 | 1.1797 |
| 4 | 3.0780 | 1.1352 |
| 5 | 3.0833 | 1.9852 |
| 5 | 3.0817 | 1.9435 |
| 6 | 3.1104 | 2.6189 |
| 6 | 3.1078 | 2.5190 |
| 7 | 3.0169 | 2.5266 |
| 7 | 3.0157 | 2.4877 |
| 8 | 2.7566 | 0.47167 |
| 8 | 2.7461 | 0.44945 |
| 9 | 2.9892 | 2.7233 |
| 9 | 2.9859 | 2.5731 |
I also examined how temperature changes the discharge behavior. At higher temperatures the electrochemical reaction is faster, but the available capacity can change because of side reactions and thermal effects. At lower temperatures the electrolyte conductivity decreases, polarization increases, and the discharge voltage drops more quickly. For an electric vehicle battery pack, this means that a forecast model trained only at room temperature may not generalize perfectly to cold or hot conditions. I therefore treated temperature as an important factor in the experimental design, even though the final model in this phase used voltage and current as the main inputs.
4. Theoretical Derivation of the Forecast Inputs
I derived the input-output structure of the forecast model from the physical relationships of the electric vehicle battery pack. Let the remaining capacity at time \(t\) be \(C(t)\), the terminal voltage be \(U(t)\), the current be \(I(t)\), the rated capacity be \(C_0\), the average internal resistance be \(R_0\), the internal resistance variation be \(R(t)\), and the delivered capacity be \(C_1(t)\). Then
$$ C(t) = C_0 – C_1(t) $$
The general functional relationship among the variables can be written as
$$ f_1\left(U(t), I(t), R(t), R_0, C(t), C_0, t\right) = 0 $$
For a specific electric vehicle battery pack, \(C_0\) and \(R_0\) are approximately constant. The internal resistance variation \(R(t)\) was small in the measured data, so I simplified the relationship to
$$ f_2\left(U(t), I(t), C(t), t\right) = 0 $$
Because the elapsed discharge time is usually unknown in an online application, I eliminated time and expressed the remaining capacity as a function of voltage and current:
$$ C(t) = g\left(U(t), I(t)\right) $$
After discretization with sampling period \(T\), the model becomes
$$ C(k) = f\left(U(k), I(k), T\right) $$
When the discharge current is constant during a sampling interval, the final form used by the neural network is
$$ C(k) = f\left(U(k), I(k)\right) $$
This result is important for the electric vehicle battery pack because it means that the remaining capacity can be inferred from terminal voltage and current without requiring a full electrochemical model. The neural network is used to approximate the nonlinear function \(f\).
| Open-Circuit Voltage (V) | Measured Internal Resistance (mΩ) |
|---|---|
| 3.65 | 22.32 |
| 3.37 | 22.46 |
| 3.082 | 22.36 |
| 2.854 | 22.37 |
| 2.564 | 22.45 |
| 2.372 | 22.43 |
| 2.005 | 22.48 |
5. BP Neural Network Model for the Electric Vehicle Battery Pack
I selected a three-layer back propagation neural network for the electric vehicle battery pack. The input layer had two nodes: discharge voltage and discharge current. The hidden layer had eleven neurons. The output layer had one node: remaining capacity. The network structure can be described as 2-11-1. I used a sigmoid activation function in the hidden layer and a linear activation function in the output layer. The mean squared error was used as the performance function:
$$ \mathrm{MSE} = \frac{1}{mp}\sum_{j=1}^{p}\sum_{i=1}^{m}\left(\hat{y}_{ji} – y_{ji}\right)^2 $$
where \(m\) is the number of output nodes, \(p\) is the number of training samples, \(\hat{y}_{ji}\) is the desired output, and \(y_{ji}\) is the network output. In my model, \(m=1\) and \(p=30\). The training samples were divided into input-output pairs. For a time series of voltage and current measurements, I grouped the data so that the network learned the mapping from voltage and current to remaining capacity.
| Input Pattern | Desired Output |
|---|---|
| \(U(1), I(1)\) | \(C(1)\) |
| \(U(2), I(2)\) | \(C(2)\) |
| \(U(3), I(3)\) | \(C(3)\) |
| … | … |
| \(U(k), I(k)\) | \(C(k)\) |
The BP algorithm adjusts weights and thresholds by gradient descent. For an output neuron, the error signal is
$$ \delta_j = (O_j – d_j)O_j(1-O_j) $$
For a hidden neuron, the error signal is
$$ \delta_j = y_j(1-y_j)\sum_{k}\delta_k w_{kj} $$
The weight update rule is
$$ w_{ij}(n+1) = w_{ij}(n) + \eta \delta_j y_i $$
where \(\eta\) is the learning rate. The threshold update follows the same principle. In my first BP model, I used a learning rate of 0.015, a maximum epoch count of 2000, and a target error of 0.01. The network was trained with the experimental data from the electric vehicle battery pack.
The first BP model produced acceptable results when the accuracy requirement was moderate. It predicted the remaining capacity of the electric vehicle battery pack with a relative error of several percent. However, the training time was long, and the model sometimes converged slowly. I also observed that the BP network can become trapped in a local minimum. This is a well-known weakness when the error surface is complex, as it is for an electric vehicle battery pack under different discharge currents and temperatures. I therefore moved to a genetic algorithm combined with BP, which I call the GA-BP method.
6. GA-BP Optimization of Network Structure Parameters
The genetic algorithm is a global search method inspired by natural selection. I used it to optimize the initial weights and thresholds of the BP network before BP training. The chromosome encoded the weights and thresholds as real numbers. The initial population size was 30. The crossover probability was 0.6, and the mutation probability was 0.2. The fitness function was based on the inverse of the squared error:
$$ f_i = \frac{1}{E_i} $$
where
$$ E_i = \sum_{p}\sum_{k}\left(T_k – V_k\right)^2 $$
Here \(T_k\) is the desired output, \(V_k\) is the network output, \(p\) is the sample index, and \(k\) is the output node index. The selection probability for individual \(i\) was
$$ P_{s,i} = \frac{f_i}{\sum_{j=1}^{N} f_j} $$
The GA-BP procedure I used can be summarized as follows. First, I initialized the population with real-valued chromosomes representing weights and thresholds. Second, I evaluated each individual using the fitness function. Third, I performed selection, crossover, and mutation. Fourth, I inserted the new individuals into the population and evaluated them again. Fifth, I repeated the genetic operations until the GA target error was reached. Sixth, I used the best chromosome as the initial weights and thresholds for BP training. Finally, I trained the BP network with the usual gradient descent until the BP target error was reached.
| GA Parameter | Value Used |
|---|---|
| Population size | 30 |
| Encoding | Real-valued |
| Crossover probability | 0.6 |
| Mutation probability | 0.2 |
| Selection | Fitness-proportionate |
The optimized initial weight matrix between the input layer and the hidden layer, \(W_1\), was
$$ W_1 = \begin{bmatrix}
-0.0090 & 0.5820 \\
0.3877 & -0.1891 \\
-0.2328 & 0.1273 \\
-0.2127 & -0.1972 \\
0.3801 & -0.6126 \\
-0.2952 & 0.8903 \\
-0.3591 & -0.0859 \\
0.2105 & -0.3088 \\
-0.6778 & 0.3019 \\
-0.5918 & -0.3165 \\
0.2864 & 0.3599
\end{bmatrix} $$
The optimized weight matrix between the hidden layer and the output layer, \(W_2\), was
$$ W_2 = \begin{bmatrix}
-0.2084 \\
-0.2235 \\
0.1563 \\
0.6178 \\
0.0618 \\
-0.6310 \\
0.2696 \\
-0.2862 \\
-0.3209 \\
0.7957 \\
-0.4142
\end{bmatrix} $$
The optimized threshold vector for the hidden layer, \(B_1\), was
$$ B_1 = \begin{bmatrix}
0.1061 \\
0.4113 \\
0.2258 \\
0.4827 \\
-0.3245 \\
-0.5378 \\
-0.6311 \\
-0.2888 \\
-0.5434 \\
-0.2097 \\
0.4065
\end{bmatrix} $$
The optimized threshold for the output layer, \(B_2\), was
$$ B_2 = 0.4126 $$
After inserting these values into the BP network, the model was retrained. The GA-BP model produced a better fit than the original BP model for the electric vehicle battery pack. The prediction curve followed the measured remaining capacity more closely, and the average difference was reduced to about 0.2 Ah in the tested range. The training time was also reduced to about 2.73 s, which was a major improvement over the initial BP training time.
| Model | Training Time (s) | Error Behavior | Remaining Capacity Fit |
|---|---|---|---|
| BP only | 22.1324 | Slow convergence, possible local minimum | Acceptable but with visible deviation |
| GA-BP with optimized initial weights and thresholds | 2.7340 | Faster convergence | Better agreement with measured values |
7. GA-BP Optimization of Training Parameters
Although optimizing the initial weights and thresholds improved the electric vehicle battery pack forecast, I also found that the training parameters strongly influence both convergence speed and accuracy. The main parameters are the target error \(goal\), the learning rate \(lr\), and the maximum number of training epochs \(epoch\). I used a nested GA procedure to optimize them. The target error was treated as the inner-loop parameter, the learning rate as the middle-loop parameter, and the maximum epoch count as the outer-loop parameter. This arrangement allowed me to examine how the three parameters interact.
First, I optimized the target error \(goal\) while keeping \(lr=0.015\) and \(epoch=2000\). The best value found was \(goal=0.02\), and the BP training time was 22.9530 s. Then I optimized both \(lr\) and \(goal\) while keeping \(epoch=2000\). The best pair was \(lr=0.013\) and \(goal=0.07\), and the training time dropped to 9.9043 s. Finally, I optimized all three parameters together. The best combination was \(epoch=1550\), \(lr=0.012\), and \(goal=0.08\), and the training time dropped to 2.7500 s.
| Optimization Stage | Parameters Optimized | Best Values | Training Time (s) |
|---|---|---|---|
| Stage 1 | goal | goal = 0.02 | 22.9530 |
| Stage 2 | lr, goal | lr = 0.013, goal = 0.07 | 9.9043 |
| Stage 3 | epoch, lr, goal | epoch = 1550, lr = 0.012, goal = 0.08 | 2.7500 |
The GA search ranges and encoding parameters were selected to balance exploration and computational cost. For the goal search, the population size was 4, the search range was 0.01 to 0.08, the termination generation was 10, the binary code length was 3, the crossover probability was 0.6, and the mutation probability was 0.01. For the learning-rate search, the population size was 8, the range was 0.001 to 0.016, the termination generation was 10, the binary code length was 4, the crossover probability was 0.6, and the mutation probability was 0.01. For the epoch search, the population size was 32, the range was 500 to 2050, the termination generation was 10, the binary code length was 5, the crossover probability was 0.6, and the mutation probability was 0.01.
| Parameter | Goal Search | Learning Rate Search | Epoch Search |
|---|---|---|---|
| Population size | 4 | 8 | 32 |
| Search range | 0.01-0.08 | 0.001-0.016 | 500-2050 |
| Termination generation | 10 | 10 | 10 |
| Binary code length | 3 | 4 | 5 |
| Crossover probability | 0.6 | 0.6 | 0.6 |
| Mutation probability | 0.01 | 0.01 | 0.01 |
After comprehensive optimization, I selected the following final training parameters for the electric vehicle battery pack forecast model: \(goal=0.08\), \(lr=0.012\), and \(epoch=1550\). The network was trained with the GA-optimized initial weights and thresholds. The training time was approximately 6.3750 s in the combined test, and the prediction accuracy remained acceptable. This showed that GA optimization can improve the speed of BP training without sacrificing the practical accuracy needed for an electric vehicle battery pack.
| Final Parameter | Value |
|---|---|
| Input nodes | 2 |
| Hidden nodes | 11 |
| Output nodes | 1 |
| Target error goal | 0.08 |
| Learning rate lr | 0.012 |
| Maximum epochs | 1550 |
| Training algorithm | GA-BP |
8. Comparative Prediction Results
I compared the BP-only model and the GA-BP model using the same discharge data from the electric vehicle battery pack. The comparison covered several voltage points at a discharge current of 3 A. The remaining capacity measured in the experiment was used as the reference. The BP model produced relative errors that were generally above 2%, while the GA-BP model reduced the relative error to approximately 1% to 3% depending on the point. This improvement is important because the remaining capacity estimate of an electric vehicle battery pack must be reliable across the flat voltage plateau, where a small voltage change can correspond to a large change in remaining capacity.
| Current (A) | Voltage (V) | Measured Remaining Capacity (Ah) | BP Prediction (Ah) | GA-BP Prediction (Ah) | BP Relative Error (%) | GA-BP Relative Error (%) |
|---|---|---|---|---|---|---|
| 3 | 3.2267 | 2.955 | 2.8110 | 2.8800 | 4.81 | 2.53 |
| 3 | 3.2132 | 2.6552 | 2.5202 | 2.7242 | 4.52 | 2.34 |
| 3 | 3.2015 | 2.3055 | 2.3355 | 2.2175 | 1.32 | 2.93 |
| 3 | 3.1917 | 2.0556 | 2.1636 | 1.9906 | 3.61 | 2.16 |
| 3 | 3.1876 | 1.9556 | 1.8146 | 2.0006 | 4.69 | 1.51 |
| 3 | 3.1757 | 1.6558 | 1.5958 | 1.5958 | 2.01 | 2.03 |
| 3 | 3.1696 | 1.5059 | 1.6259 | 1.5899 | 4.05 | 2.83 |
I also tested the generalization of the electric vehicle battery pack forecast model by predicting values that were not directly sampled. The sampling interval of the test equipment was fixed, so many intermediate points were not measured. I selected a discharge current of 3 A and voltage points from 3.30 V to 3.10 V in steps of 0.02 V. The model produced the following remaining-capacity estimates. These points lie on the flat part of the discharge curve, where accurate estimation is especially difficult.
| Current (A) | Voltage (V) | Predicted Remaining Capacity (Ah) |
|---|---|---|
| 3 | 3.30 | 2.9735 |
| 3 | 3.28 | 2.9682 |
| 3 | 3.26 | 2.8653 |
| 3 | 3.24 | 2.7192 |
| 3 | 3.22 | 2.5634 |
| 3 | 3.20 | 2.3436 |
| 3 | 3.18 | 1.9438 |
| 3 | 3.16 | 1.7466 |
| 3 | 3.14 | 1.5968 |
| 3 | 3.12 | 1.3569 |
| 3 | 3.10 | 1.2442 |
To test generalization further, I used a discharge current of 2.5 A, which was not part of the original training set. The voltage range was again chosen from 3.40 V to 3.20 V. The forecast results are listed below. This test is important because it represents an operating condition that the electric vehicle battery pack may encounter but that was not explicitly used during training.
| Current (A) | Voltage (V) | Predicted Remaining Capacity (Ah) |
|---|---|---|
| 2.5 | 3.40 | 2.9846 |
| 2.5 | 3.38 | 2.9666 |
| 2.5 | 3.36 | 2.8935 |
| 2.5 | 3.34 | 2.7752 |
| 2.5 | 3.32 | 2.6024 |
| 2.5 | 3.30 | 2.4368 |
| 2.5 | 3.28 | 2.0456 |
| 2.5 | 3.26 | 1.8985 |
| 2.5 | 3.24 | 1.7123 |
| 2.5 | 3.22 | 1.4650 |
| 2.5 | 3.20 | 1.3368 |
9. Experimental System and Accuracy Validation
I then built an experimental system around the trained GA-BP model. The system accepted current and voltage as inputs and returned the estimated remaining capacity of the electric vehicle battery pack. To validate the system, I adjusted the sampling interval to 10 s and performed independent discharges at 2.5 A and 3 A. The measured remaining capacities were compared with the predicted values. This validation was more demanding than the earlier tests because the data points were selected independently and the sampling resolution was finer.
For the 3 A discharge, the predicted and measured values were close across the tested voltage range. The relative error was approximately 1% in most cases, with a maximum near 1.73%. This level of accuracy is acceptable for many engineering uses of an electric vehicle battery pack, especially when the estimate is used for range display, energy management, and diagnostic trend analysis.
| Current (A) | Voltage (V) | Predicted Capacity (Ah) | Measured Capacity (Ah) | Relative Error (%) |
|---|---|---|---|---|
| 3 | 3.40 | 2.9735 | 2.9315 | 1.39 |
| 3 | 3.38 | 2.9682 | 2.9552 | 0.43 |
| 3 | 3.36 | 2.8653 | 2.8893 | 0.80 |
| 3 | 3.34 | 2.7192 | 2.6862 | 1.09 |
| 3 | 3.32 | 2.5634 | 2.5114 | 1.73 |
| 3 | 3.30 | 2.3436 | 2.3836 | 1.33 |
| 3 | 3.28 | 1.9438 | 1.9098 | 1.13 |
| 3 | 3.26 | 1.7466 | 1.7676 | 0.69 |
| 3 | 3.24 | 1.5968 | 1.5648 | 1.07 |
| 3 | 3.22 | 1.3569 | 1.3309 | 0.87 |
| 3 | 3.20 | 1.2442 | 1.2142 | 1.00 |
For the 2.5 A discharge, the relative error was slightly higher, generally between 1.6% and 2.9%. This result is reasonable because 2.5 A was not included in the training data, so the test reflects the interpolation ability of the model. The error remained within an acceptable engineering range for the electric vehicle battery pack. The results suggest that the GA-BP model captures the dominant nonlinear relationship between voltage, current, and remaining capacity.
| Current (A) | Voltage (V) | Predicted Capacity (Ah) | Measured Capacity (Ah) | Relative Error (%) |
|---|---|---|---|---|
| 2.5 | 3.40 | 2.9846 | 2.8966 | 2.93 |
| 2.5 | 3.38 | 2.9666 | 2.8906 | 2.53 |
| 2.5 | 3.36 | 2.8935 | 2.9545 | 2.03 |
| 2.5 | 3.34 | 2.7752 | 2.7062 | 2.29 |
| 2.5 | 3.32 | 2.6024 | 2.6544 | 1.73 |
| 2.5 | 3.30 | 2.4368 | 2.3708 | 2.21 |
| 2.5 | 3.28 | 2.0456 | 1.9686 | 2.56 |
| 2.5 | 3.26 | 1.8985 | 1.9765 | 2.62 |
| 2.5 | 3.24 | 1.7123 | 1.6453 | 2.23 |
| 2.5 | 3.22 | 1.4650 | 1.4130 | 1.73 |
| 2.5 | 3.20 | 1.3368 | 1.2888 | 1.60 |
10. Discussion of the GA-BP Forecast for Electric Vehicle Battery Packs
The results show that a neural-network model can forecast the remaining capacity of an electric vehicle battery pack without a detailed electrochemical model. The BP network alone is capable of learning the input-output mapping, but it suffers from slow convergence and local minima. The GA-BP method improves the situation by searching for good initial weights and thresholds before BP training. It also allows the training parameters to be tuned in a systematic way. For an electric vehicle battery pack, this matters because the battery management system must produce an estimate quickly and repeatedly during driving.
The flat voltage plateau of lithium iron phosphate is both an advantage and a challenge. It provides stable operation, but it makes voltage-based state estimation difficult because a small voltage change can correspond to a large change in remaining capacity. The neural network is well suited to this problem because it can use current as an additional input and can learn the subtle shape of the discharge curve. The GA-BP model performed well in the plateau region in my tests, although some error remained. The error is influenced by temperature, aging, self-discharge, and the internal resistance variation of the electric vehicle battery pack.
From an engineering perspective, the model has several practical benefits. It does not require a long rest period, unlike open-circuit voltage methods. It does not accumulate error in the same way as pure ampere-hour integration because it uses voltage as a correcting variable. It can be implemented in a microcontroller or an embedded controller if the network is kept small. The 2-11-1 structure I used is compact enough for online use in an electric vehicle battery pack. The use of GA optimization can be performed offline, so the online computational burden remains low.
However, I also recognize the limitations. The model was trained on a specific lithium iron phosphate cell type and a limited range of currents and temperatures. An electric vehicle battery pack contains many cells, and cell-to-cell inconsistency can affect the pack-level voltage and capacity. The electric vehicle battery pack also experiences regenerative charging, dynamic current profiles, and temperature gradients that are not fully represented in constant-current discharge tests. Therefore, the model should be extended with more inputs, such as temperature, internal resistance, and aging indicators, before it is applied to a full vehicle battery management system.
| Factor | Influence on Remaining Capacity Forecast | Possible Improvement |
|---|---|---|
| Temperature | Changes reaction rate, polarization, and available capacity | Add temperature as a network input |
| Internal resistance | Reflects aging and power capability | Estimate online and include as input |
| Self-discharge | Reduces stored charge during rest | Use periodic correction and history data |
| Cell inconsistency | Weak cells limit pack performance | Use per-cell or group-level features |
| Dynamic current | Creates transient voltage and recovery effects | Train with real driving profiles |
11. Implementation Details and Mathematical Summary
For the electric vehicle battery pack forecast, I used the following overall procedure. First, I collected voltage and current data during constant-current discharge. Second, I computed the delivered capacity by integrating current over time. Third, I derived the remaining capacity from the rated capacity and delivered capacity. Fourth, I normalized the data for neural-network training. Fifth, I designed a 2-11-1 BP network. Sixth, I used a genetic algorithm to optimize the initial weights and thresholds. Seventh, I optimized the training parameters with a nested GA search. Eighth, I trained the final GA-BP network and validated it on independent data. Ninth, I compared the predicted remaining capacity with measured values and evaluated the relative error.
The normalized input can be written as
$$ x_i’ = \frac{x_i – x_{\min}}{x_{\max} – x_{\min}} $$
The network output is then denormalized as
$$ y = y_{\min} + y’\left(y_{\max} – y_{\min}\right) $$
The sigmoid activation used in the hidden layer is
$$ \phi(v) = \frac{1}{1 + e^{-v}} $$
The forward propagation for the hidden layer can be written as
$$ h_j = \phi\left(\sum_{i=1}^{2} w_{ji}^{(1)} x_i + b_j^{(1)}\right) $$
The output layer is
$$ \hat{C} = \sum_{j=1}^{11} w_j^{(2)} h_j + b^{(2)} $$
where \(\hat{C}\) is the predicted remaining capacity of the electric vehicle battery pack. The training error is
$$ E = \frac{1}{2}\sum_{k=1}^{N}\left(C_k – \hat{C}_k\right)^2 $$
The BP weight update is
$$ w_{ij}^{(l)}(n+1) = w_{ij}^{(l)}(n) – \eta \frac{\partial E}{\partial w_{ij}^{(l)}} $$
The GA fitness is
$$ F = \frac{1}{E + \epsilon} $$
where \(\epsilon\) is a small positive number to avoid division by zero. The selection probability is
$$ P_i = \frac{F_i}{\sum_{j=1}^{N_p} F_j} $$
These equations summarize the mathematical foundation of my electric vehicle battery pack forecast system. The combination of BP and GA gives the model both local search capability and global search capability. The BP algorithm refines the weights, while the GA avoids poor initial conditions and helps the network escape local minima.
| Symbol | Meaning |
|---|---|
| \(C_{\text{rem}}\) | Remaining capacity of the electric vehicle battery pack |
| \(C_{\text{rated}}\) | Rated capacity |
| \(U\) | Terminal voltage |
| \(I\) | Discharge current |
| \(R_0\) | Internal resistance |
| \(T\) | Sampling period |
| \(\eta\) | Learning rate |
| \(w_{ij}\) | Connection weight |
| \(b_j\) | Threshold or bias |
| \(F\) | GA fitness value |
12. Main Findings and Engineering Meaning
My main finding is that the remaining capacity of an electric vehicle battery pack can be forecast with acceptable accuracy by using a GA-BP neural network that takes voltage and current as inputs. The method avoids the need for a full electrochemical model and does not require the battery to rest before measurement. The GA-BP model outperformed the BP-only model in both convergence speed and prediction accuracy. The optimized initial weights and thresholds reduced the training time substantially, and the optimized training parameters further improved efficiency.
For the electric vehicle battery pack, the practical implications are significant. A faster training process means that the model can be updated more often with new data. A more accurate forecast means that the vehicle controller can make better decisions about available range, power limits, and charging strategy. The model can also be used in a laboratory test system to predict remaining capacity at points that were not directly measured. This reduces the time and cost of exhaustive discharge testing.
The validation results showed relative errors generally within about 1% to 3% for the tested conditions. This is within an acceptable engineering range for many battery management tasks. However, I do not claim that the model is universal. It was trained for a specific lithium iron phosphate cell and a specific set of currents and temperatures. A production electric vehicle battery pack would require additional data and possibly additional inputs. Temperature, aging, and cell imbalance are the most important factors to include in future work.
| Achievement | Result |
|---|---|
| Input-output model | Voltage and current as inputs, remaining capacity as output |
| Base model | 2-11-1 BP neural network |
| Optimization method | GA-BP with optimized initial weights, thresholds, and training parameters |
| Accuracy | About 1% to 3% relative error in validation tests |
| Efficiency | Training time reduced from tens of seconds to a few seconds |
| Application target | Online estimation for an electric vehicle battery pack |
13. Future Improvements
To make the electric vehicle battery pack forecast more robust, I would extend the input vector to include temperature and internal resistance. Temperature can be measured with sensors in the pack, and internal resistance can be estimated from voltage and current transients. I would also train the network with dynamic current profiles that resemble real driving, including acceleration, regenerative braking, and rest periods. This would help the model learn the recovery effect and the transient behavior of the electric vehicle battery pack.
Another improvement is to use a recurrent neural network or a nonlinear autoregressive model with exogenous inputs. Such a model can use previous voltage, current, and capacity estimates as additional inputs. This may capture history-dependent effects more effectively than a static feedforward network. I would also consider a hybrid model that combines ampere-hour integration with the neural forecast, so that the long-term drift of integration is corrected by the voltage-based neural estimate. For an electric vehicle battery pack, this hybrid approach may provide a good balance between short-term response and long-term stability.
Finally, I would test the method on a full electric vehicle battery pack rather than a single cell. Pack-level behavior includes unequal cell aging, thermal gradients, and parallel path currents. The forecast system would need to aggregate cell-level features or use a reduced-order pack model. With these extensions, the GA-BP method could become a practical component of a battery management system for an electric vehicle battery pack.
14. Conclusion
In my study of remaining capacity forecasting for an electric vehicle battery pack, I combined experimental battery characterization, nonlinear modeling, and intelligent optimization. I found that voltage and current are sufficient to build a useful neural-network forecast for the tested lithium iron phosphate cell. The BP network alone can learn the mapping, but its training is slow and its error surface can trap the solution in a local minimum. By using a genetic algorithm to optimize the initial weights and thresholds, I obtained a better starting point for BP training. By further optimizing the target error, learning rate, and maximum epoch count, I reduced the training time while maintaining accuracy.
The validated system predicted remaining capacity with relative errors generally within 1% to 3% for the tested electric vehicle battery pack conditions. This satisfies the practical requirement for many engineering applications and confirms that the GA-BP method is reliable and feasible. The approach is especially attractive because it does not require a complex electrochemical model and can be trained offline with measured data. It can then be embedded in an online estimator for an electric vehicle battery pack.
The main limitation is that the model must be extended to include temperature, aging, self-discharge, and cell inconsistency before it can be applied to all operating conditions of a real electric vehicle battery pack. Nevertheless, the results demonstrate that intelligent forecasting is a valuable direction for electric vehicle battery pack management. The combination of BP and GA provides a foundation for accurate, efficient, and practical remaining capacity prediction.